AI may soon outsmart humans
4 min read
To have any hope of regulating AI effectively, legislators need to have a basic understanding of how current AI works.
Actually, AI already scores better on standard tests of creativity than most people. Rather than speculating about how to regulate AI, I think it will be more useful to explain how it works.
For more than 50 years, the great majority of AI researchers were convinced that AI must work like logic. Given some true premises and some rules of inference, logic can derive new facts by applying the rules of inference to generate new strings of symbols. This paradigm failed.
Once you accept that LLMs understand sentences in much the same way as we do, it is obvious that they really do understand
A quite different and widely derided approach called neural networks was inspired by biology. Instead of treating logical reasoning as the essence of intelligence, this approach viewed perception, motor control and analogical reasoning as the core skills of biological intelligence and was committed to the belief that these skills could be learnt by adapting the connection strengths between simple neuron-like processing units.
With sufficient training data and computing power, the neural network paradigm finally won out and led to the large language models (LLMs) and image generators that are currently getting so much attention. But how do these LLMs work, and do they really understand what they are saying?
When an LLM processes text, it first divides it into small word fragments like “un” or “ing” or “may” and for each fragment it creates an initial “embedding” which is just a set of activity levels for some features. One feature might have a high activity level if the fragment has something to do with animate objects, and another feature might have a high activity level if the fragment has something to do with time.
The embedding is intended to capture the meaning of the fragment, but this cannot be done by just looking at that fragment – it depends on the context. For example, the fragment “may” could be the month, a woman’s name, or a modal, as in: “you may think that”. So the initial embedding has to be refined. If the nearby text contains the fragment “feb”, for example, the embedding for “may” can be refined by raising the activity level of the temporal feature and lowering the activity level of the animate feature. The interactions between features eventually lead to a consistent set of embeddings for all the word fragments that capture the meaning of the sentence and make it possible to predict what comes next.
The raw symbols that initially represent the fragments act like a skeleton which then gets fleshed out by all the interactions between the active features in the embeddings. That is how sentences get meaning.
It involves about a trillion learnt weights that determine how the features of different embeddings should interact, which means there will never be any simple explanation of why an LLM interprets a particular sentence in a particular way. Something similar probably goes on in our brains. Like the LLMs, we do not construct meanings for sentences by using set rules that operate on symbol strings. We use trillions of interactions between billions of features that are represented by the activity levels of the neurons in our brains.
Once you accept that LLMs understand sentences in much the same way as we do, it is obvious that they really do understand. It is true that they are trained just to predict the next word, but to do this really well they have to understand what was said.
LLMs construct an immensely complicated model of their training data that uses interactions between features to capture all the structure in the data. This kind of model is understanding.
We have created a digital form of intelligence that really does understand, and it may soon be smarter than the biological intelligence that created it.
Geoffrey Hinton, AI pioneer and former Google vice-president
PoliticsHome Newsletters
Get the inside track on what MPs and Peers are talking about. Sign up to The House's morning email for the latest insight and reaction from Parliamentarians, policy-makers and organisations.
